This repo contains the code and source docs for our paper:
Create a virtual environment:
cd /path/to/moderl
python -m venv moderl-venv
source moderl-venv/bin/activate
Install ModeRL
in editable mode with dependencies needed for experiments:
pip install -e ".[experiments]"
See experiments/ for detailed instructions on running all of the experiments in the paper.
As an example, the ModeRL
experiment with a schedule that tightens the constraint level during training can be run with:
cd ./experiments
python train.py +experiment=constraint_schedule
See the example notebook to see how to use ModeRL
in practice.
I use git subtrees to manage dependencies on my own packages. In particular, my Mixtures of Sparse Variational Gaussian Process Experts package mosvgpe and my simulation environments package simenvs.
If I make changes to https://github.com/aidanscannell/mosvgpe I can pull them using,
git subtree pull --prefix=subtrees/mosvgpe mosvgpe-subtree master
And when I make changes to mosvgpe
in moderl
I can push the changes back
to https://github.com/aidanscannell/mosvgpe using,
git subtree push --prefix=subtrees/mosvgpe mosvgpe-subtree /branch/to/push/to
For example,
git subtree push --prefix=subtrees/mosvgpe mosvgpe-subtree aidanscannell/push-from-moderl
@proceedings{scannell2023moderl,
title={Mode-constrained Model-based Reinforcement Learning via Gaussian Processes},
author={Scannell, Aidan and Ek, Carl Henrik and Richards, Arthur},
booktitle = {International {{Conference}} on {{Artificial Intelligence}} and {{Statistics}}},
year={2023}
}